Summary of the invention
(1) technical matters that will solve
In view of this, fundamental purpose of the present invention provides a kind of iris classification method based on texture primitive statistical characteristic analysis, to accelerate iris recognition technology carries out the feature templates comparison in large scale database speed.
(2) technical scheme
For achieving the above object, the invention provides a kind of iris classification method based on texture primitive statistical characteristic analysis, this method comprises:
S1, in the training set clearly iris image carry out pre-service, obtain region of interest ROI, feature extraction is carried out in the ROI zone, the textural characteristics that extracts is trained and modeling, obtain iris rough sort model;
S2, the clear iris image of any input is carried out pre-service, obtain the ROI zone, carry out feature extraction then, the iris texture characteristic that extraction is obtained is input in the model that step S1 training obtains, and obtains the classification information of input iris image.
In the such scheme, described step S1 comprises:
S11, to the input iris image carry out pre-service;
S12, the normalized iris image in the training set is carried out texture analysis, extract the textural characteristics that obtains each pixel and neighborhood thereof;
S13, the textural characteristics that obtains is carried out cluster, obtain N classification;
S14, to each width of cloth iris image, with the texton histogram corresponding that obtain overall textural characteristics as this image with it;
After the texton histogram of all iris images in S15, the extraction training set, utilize clustering method once more, iris image is divided into the M class, obtain iris rough sort model.
In the such scheme, described in the step S1 iris image of importing is carried out pre-service and comprise Iris Location and normalization, specifically comprise: at first the gray level image to input carries out iris detection and cuts apart, use the mode conversion of bilinearity difference under polar coordinates the iris image under the cartesian coordinate system then, polar initial point is exactly the center of circle of pupil, under polar coordinate system, all iris images are zoomed to unified size, realize the normalization of iris image.
In the such scheme, described step S2 comprises:
S21, the iris image of current input is carried out pre-service, obtain the normalization iris image;
S22, normalization iris image to obtaining, the texture primitive according to training among the step S13 obtains calculates its texton histogram;
S23, be written into the model that obtains among the step S15, and, calculate the distance at input feature vector and each big class center respectively texton histogram feature input rough sort module;
The category label of S24, selected distance minimum is finished the rough sort process as the category label of this input picture.
(3) beneficial effect
From technique scheme as can be seen, the present invention has following beneficial effect:
1, this iris classification method provided by the invention based on texture primitive statistical characteristic analysis, by extracting the textural characteristics of iris image, iris image is divided into several classifications, and the iris feature template in the database is just arranged according to the classification under its image.When carrying out the iris comparison, at first carry out the rough sort of iris, obtain the classification information of input picture, and then belong to of a sort feature templates with input picture with it and compare, thereby make shorten the averaging time of finishing an iris comparison, reach real-time effect, accelerated iris recognition technology carries out the feature templates comparison in large scale database speed effectively.
2, this iris classification method based on texture primitive statistical characteristic analysis provided by the invention can effectively improve the real-time of extensive iris authentication system.At first judge the classification of iris, only in this classification, search for current user's identity then, can reduce the search volume of algorithm, carry out the required time of iris feature comparison thereby reduce greatly.
3, this iris classification method based on texture primitive statistical characteristic analysis provided by the invention can improve the accuracy of extensive iris authentication system.At first interpretation goes out the classification of iris, has been equivalent to reduce the scale of iris recognition problem, has also just reduced the iris recognition probability of errors, thereby improves the accuracy of Algorithm of Iris Recognition.
4, this iris classification method based on texture primitive statistical characteristic analysis provided by the invention has been studied the relation between iris texture and the gene genetic, has studied the user with similar gene and whether has had similar iris texture image.
Embodiment
Describe each related detailed problem in the technical solution of the present invention in detail below in conjunction with accompanying drawing.Be to be noted that described embodiment only is intended to be convenient to the understanding of the present invention, and it is not played any qualification effect.
In existing iris authentication system, the input iris image must be compared just one by one with all iris images in the database can draw last comparison result, though generally being the very fast Hamming distance of employing speed, the iris comparison compares mode, but continuous increase along with user number in the database, once compare the also constantly increase of needed time, in ultra-large iris database, can't satisfy the demand of system real-time.The present invention is divided into several classifications by extracting the textural characteristics of iris image with iris image, and the iris feature template in the database is just arranged according to the classification under its image.When carrying out the iris comparison, at first carry out the rough sort of iris, obtain the classification information of input picture, and then belong to of a sort feature templates with input picture with it and compare, thereby make shorten the averaging time of finishing an iris comparison, reach real-time effect.
Iris rough sort method based on texture analysis plays a very important role for the real-time and the accuracy tool that improve extensive iris authentication system.Utilize study to obtain the method for texture primitive, the present invention has realized one based on the texture primitive statistical characteristic analysis iris classification method.Fig. 1 is the process flow diagram based on the texture primitive statistical characteristic analysis iris classification method provided by the invention, comprises the training and two parts of classifying:
Described training process comprises step S1: in the training set clearly iris image carry out pre-service, obtain area-of-interest---the ROI zone, feature extraction is carried out in the ROI zone, the textural characteristics that extracts is trained and modeling, obtain iris rough sort model.
In this step, it is the local grain feature of extracting the ROI zone that feature extraction is carried out in the ROI zone, obtain the iris texture primitive after the extraction, obtain iris texture primitive histogram after the local grain feature is shone upon according to the iris texture primitive, adopt clustering method that iris image is divided into N subclass.The textural characteristics of described extraction iris image is also trained, and obtains the rough sort that the iris disaggregated model is used for iris image.Described training is the least unit that obtains characterizing iris texture by self-defined or learning process---an iris texture primitive, then by making up the textural characteristics of iris texture primitive histogram as iris image.
Described assorting process comprises step S2: the clear iris image to any input is carried out pre-service, obtain the ROI zone, carry out feature extraction then, the iris texture characteristic that extraction is obtained is input in the model that step S1 training obtains, and obtains the classification information of input iris image.
In this step, in the iris classification, be to use the method calculating input image of histogram coupling and the distance between the class models to obtain classified information.Iris classification can be regarded the thick coupling (coarse-level matching) of iris as, and it can constitute an accurately iris authentication system fast with other iris identification method.
Above-mentioned training step S1 specifically comprises the steps:
Step S11: the iris image to input carries out pre-service.Promptly at first the gray level image of importing is carried out iris detection and cuts apart, use the mode conversion of bilinearity difference under polar coordinates the iris image under the cartesian coordinate system then, polar initial point is exactly the center of circle of pupil, under polar coordinate system, all iris images are zoomed to unified size, be called the normalization iris image;
Step S12: the normalized iris image in the training set is carried out texture analysis, each pixel and neighborhood thereof are carried out the textural characteristics extraction, each such textural characteristics has just been expressed near the texture features current this pixel.
Step S13: the textural characteristics among the step S12 is carried out cluster, obtains N classification, the proper vector of each cluster centre represented a kind of in iris image the higher textured pattern of the frequency of occurrences, be called as texture primitive.
Step S14: to each width of cloth iris image, with the texton histogram corresponding that obtain overall textural characteristics as this image with it.
Step S15: according to step S14, after all iris images extraction texton histogram in the training set, utilize clustering method once more, iris image is divided into the M class.The average texture primitive histogram of each class is exactly such other average expression, the model parameter that training just obtains.
Above-mentioned classification step S2 specifically comprises the steps:
Step S21: the iris image to current input carries out pre-service, obtains the normalization iris image.
Step S22: to the normalization iris image that step S21 obtains, the texture primitive according to training among the step S13 obtains calculates its texton histogram.
Step S23: be written into the model that obtains among the step S15, and, calculate the distance at input feature vector and each big class center respectively with texton histogram feature input rough sort module.
Step S24: use the distance that obtains among the step S23, the category label of selected distance minimum is finished the rough sort process as the category label of this input picture.
The committed step that method of the present invention is related to describes in detail one by one below, and the basic step in the method for the present invention is identical, and concrete form is as described below:
The first, be the pre-service of iris image.Not only comprise iris in the iris image, also have pupil, sclera, eyelid and eyelashes etc.Therefore will carry out the iris classification, the first step should be from iris image iris to be separated, and then the iris annulus is normalized to the rectangular area of fixed measure, i.e. the pre-service of iris (Iris Location and normalization), and this is a critical step.
1, Iris Location
The all very approaching circle of the pupil of human eye and iris outline, so we adopt round model to come match pupil and iris boundary.The gray scale of human eye pupil is lower than the peripheral region, so can use threshold method to be partitioned into pupil region, center of gravity that then should the zone is gone to the edge of match pupil as preliminary pupil center with the template of variable dimension near this point, best fitting result is exactly the positioning result of pupil.The center of iris is near the center of pupil, so the center and the radius that can use the same method and find iris.Because human eye is an internal of human body, iris is easy to be blocked by the eyelid eyelashes, and in order to improve the reliability of entire method, we have chosen the iris region that is blocked least easily as our interesting areas (ROI).Fig. 2 (b) is to the example after the Iris Location among Fig. 2 (a), and wherein the solid white line circle is represented the outer boundary of pupil and iris after the match, and dashed rectangle is represented area-of-interest
2, normalization
In the mode of bilinearity difference, the iris annulus of having good positioning can be carried out the rectangular area of spatial alternation to a fixed measure.Fig. 2 (c) is the result after the iris normalization, and dashed rectangle has been represented the position of ROI zone in normalized image.
Each width of cloth iris image clearly can both obtain size and is 256 * 60 ROI zone and carry out following feature extraction through after the iris pre-service.
The second, be the acquisition of iris texture primitive in the training process.
Iris image can be regarded a kind of texture in a sense as and distribute, and this texture is to be made of various iris texture primitives with certain specific character.The residing position of these texture primitives has nothing in common with each other in the iris image of different eyes, and this makes iris become a kind of mode of biological characteristic very accurately.But the number of various texture primitives and distribution situation are similar in some iris image, make these iris images from visually seeming very alike.Therefore we are from intuitively thinking that these alike iris images should be divided into same class.
Carry out the iris classification, we at first need to define the iris texture primitive.We have two kinds of methods to obtain texture primitive in the present invention: the firstth, and self-defined texture primitive, the secondth, by the method acquisition texture primitive of machine learning.
So-called self-defined texture primitive is exactly that artificial image pixel that will satisfy certain particular kind of relationship and neighborhood thereof is defined as a kind of texture primitive.For example LBP (local binaryzation pattern, LocalBinary Pattern) is exactly a kind of self-defining texture primitive.
In example of the present invention, we adopt the method for machine learning to obtain texture primitive.As shown in Figure 3, for the ROI zone that we obtain in the training storehouse, we carry out filtering with one group of wave filter to it earlier, and the filtering result of each pixel can represent with a proper vector.Then these all proper vectors are sent into machine learning algorithm program (we adopt the K mean algorithm) here and carried out cluster, obtain N (N=64 in this example) cluster centre, each cluster centre is represented a kind of texture primitive.
Three, be the histogrammic calculating of iris texture primitive.
For the ROI zone of each width of cloth iris image, obtain a series of filtering results after our the process bank of filters filtering.The filtering result of each pixel is exactly the proper vector of a local grain, calculates the Euclidean distance of this vector to each texture primitive, gets the mapping result of a wherein minimum texture primitive as current this pixel.
All pixels in each ROI zone are through obtaining the histogram of iris texture primitive by mapping after the filtering like this.Though single iris texture primitive is expressed is a pixel and the local grain information in the neighborhood on every side thereof, but iris texture primitive histogram has also been expressed the overall texture information of a width of cloth iris image, and classification is a kind of very effective information for iris for this.
Four, be on training set, to calculate the iris disaggregated model.
For all the ROI zones in the training set, we can obtain its corresponding iris texture primitive histogram as its textural characteristics.In the present invention, we adopt card side's distance to weigh two similarity degrees between the texton histogram.The concrete formula of the side's of card distance is as follows:
Wherein H1 and H2 represent two iris texture primitive histograms respectively.Because H1
i+ H2
iMay equal zero, so we only consider nonzero term.
After having defined the distance calculation formula, we will adopt the method for machine learning to obtain the parameter of iris disaggregated model once more, we adopt the training of K mean cluster method to obtain the model parameter of five class iris images in this example, and wherein the average texture primitive histogram of each class iris image is exactly such other model parameter.Fig. 5 has provided the typical image in each classification in the last classification results.In the iris categorizing system of reality,, can not effectively shorten and finish one time the required time of iris recognition if the classification of classification very little; If class categories is too many, be difficult to guarantee the accuracy of iris classification again.After balance, in this example, we select K=5.
Five, be to carry out the iris classification.
In application process, for the iris image of any width of cloth input, by top step, we are not difficult to obtain its corresponding iris texture primitive histogram.We are in the disaggregated model that obtains in this histogram substitution training process, calculate the similarity of this histogram and each classification, and the line ordering of going forward side by side is used the category label mark input picture of similarity maximum at last.
In ensuing iris comparison process, we with input picture at first and the feature templates that belongs to same classification with it compare, thereby make shorten the averaging time of finishing an iris comparison.
For verification algorithm validity, use the CASIA iris database that the algorithm that proposes is tested.The CASIA iris database is a shared data bank of being created by Institute of Automation, CAS, is used to evaluate and test Algorithm of Iris Recognition, at present by how tame research unit employing in the world.The CASIA iris database comprises the iris image of 800 eyes.According to test of heuristics, correct classification rate is 95.0%.And the accuracy rate that adds the iris authentication system of iris sorting algorithm also improves, its etc. error rate (EER) be reduced to 0.88%. from 1.1%
Does so great database use method of the present invention could improve the speed of system? we suppose that the size of iris database is N, T1 represents to extract the time of carrying out the required feature of iris recognition, T2 represents to extract and carries out iris and classify time of required feature, T3 represents to carry out the time of an iris comparison, and T4 represents to carry out the time of iris classification.In this example, the iris classification speed is very fast, can think T4=0, and other times are respectively T1=45ms, if T2=660ms and T3=1.1ms. iris classification required time T2 mate the time of saving less than iris after adopting new method, just can think that it is effective carrying out the iris classification earlier.
When not adopting iris to divide time-like, be the averaging time of finishing iris coupling:
T
without=T
1+0.5*N*T
3 (2)
When employing after to be divided into five classes and correct classification rate be 95.0% iris sorting algorithm, be the averaging time of finishing an iris coupling:
T
with=T
1+T
2+(95.0%*0.2+5%*1)*0.5*N*T
3(3)
Make T
Without=T
WithWe are not difficult to obtain number that N=1579. just works as registered iris feature template in the database greater than 1579 the time, adopt method of the present invention can improve the average behavior of iris authentication system.Suppose to have in the iris database 10,000 iris feature templates, we are not difficult to obtain and adopt this method will save iris match time about 63%, and along with the increase of database scale, the time of saving will be more considerable.
Specific embodiment provided by the invention is as follows:
The present invention is particularly suitable for having the iris authentication system of extensive iris database.What adopt as the self-service boarding system on certain airport is iris recognition technology, and the fugitive suspect's of this system and public security bureau iris database (scale probably has 1,000,000) links to each other.Because embezzlement is ordered to arrest by public security organ, his iris information just has been put in fugitive suspect's the iris database (being also referred to as blacklist) as Zhang San.When Zhang San disguises oneself, use counterfeit passport to prepare to run away by air, by plane, when using self-service boarding system, Zhang San's iris image has been taken by system, automatically extract its classification information, at last iris image enrollment generic in this iris image and the blacklist is compared one by one, confirmed Zhang San's true identity, self-service boarding system begins to report to the police, and whole process was just finished within 5 seconds.Though Zhang San has forged certificate, he is still arrested and is brought to justice under help of the present invention.
The present invention can be the gordian technique in the iris authentication system of future generation at the overall performance that effectively improves iris authentication system aspect recognition speed and the accuracy rate.
The above; only be the embodiment among the present invention; but protection scope of the present invention is not limited thereto; anyly be familiar with the people of this technology in the disclosed technical scope of the present invention; can understand conversion or the replacement expected; all should be encompassed in of the present invention comprising within the scope, therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.